A Human-Centred Analytics Framework for Auditable Supply Chain Risk Monitoring and Decision Support

Supply chain managers must monitor multi-domain logistics data continuously, detect emergent risks and act faster than manual analysis allows. Large Language Models (LLMs) reason contextually but carry latency and hallucination risk unsuitable for continuous detection. Three gaps persist: AI-enabled supply chain risk monitoring remains predominantly retrospective; no framework implements hybrid LLM-SLM (Small Language Model) deployment in multi-agent architectures or combines rule-based detection with practitioner-facing natural language decision support. This paper addresses these gaps through a human-centred, hierarchical three-tier multi-agent framework that separates detection from explanation: deterministic rules handle rule-violation detection and risk scoring; SLM worker agents, invoked only when rule violations are confirmed, generate domain-specific summaries; and a supervisor LLM synthesises cross-domain findings into a structured risk report. The human operator retains full decision authority at the third tier. Because detection is deterministic and rule-based, it is fast and verifiable, and hallucination is excluded from the detection layer by construction; residual hallucination risk is confined to the generative explanation layer, where human review remains necessary before any operational action. Evaluated on a 32,065-record logistics dataset, the framework screens eight SLM architectures (1 to 7 billion parameters) and benchmarks four finalists on 100 held-out records. The preferred candidate, Granite 3.1 MoE 3B, reaches 112.2 tokens per second on edge hardware (6GB VRAM). Compression loss, the omission of items carried by the reference report, accounts for 86.6 per cent of classified divergences. At scale (1,000 records), hybrid edge-cloud configuration reduces processing time by 57 per cent and API cost by 39.6 per cent versus fully cloud-based deployment. The evaluation is batch-mode proof of concept; live-stream validation and practitioner user study remain future work. The paper contributes to data analytics through rule-based detection, composite risk scoring, and multi-model benchmarking; and a hybrid edge-cloud AI architecture; and to decision support across supply chain.